replay: overlap execution and improve checkpoints and epoch recovery - #292
Draft
7layermagik wants to merge 116 commits into
Draft
7layermagik wants to merge 116 commits into
7layermagik wants to merge 116 commits into
Conversation
Share lazy encoding state through capture and pruning without retaining parent links or copying synchronization primitives. Keep MTS2 bytes and caller-owned output unchanged; allocate exactly sized node and checkpoint buffers. Check the original wire encoder, concurrent pruning and encoding, output ownership and recovery. Benchmark moving windows including the default 128-root cadence; document retained-memory costs and the all-new-window limit.
Record moving-window results at the real 128-root cadence, retained-memory tradeoffs and unchanged cold-window costs. Keep raw benchmark artifacts outside the source tree; do not infer live voting gains from staging measurements.
…en transfer) Consensus-neutral performance changes to pkg/sbpf, validated against the unmodified interpreter with a 100k-program differential corpus (identical return values, errors/PCs, CU consumed, meter remaining, memory contents, input-region state) plus the package's unit tests: - meter instructions with a local due/budget pair synced around syscalls and on exit (Agave's due_insn_count scheme) instead of calling ComputeMeter.Consume per instruction - move cold opcodes to executeCold so Run drops below the compiler's "big function" threshold and Consume/Read*/Push/Pop/fast paths inline - zero only the dirty range of the pooled stack/heap in Finish (page bitmap on the fast path, byte range on the translate path) instead of 256 KiB + HeapMax per execution - per-window fast-path address translation table (Agave aligned mapping layout, branch-free v0 frame gaps, one-entry cache for VASA input regions) - 16-wide register file (no bounds checks on r[dst]/r[src]), in-place call-frame Push/Pop, precomputed internal call targets per Program pooling_test writes through the VM's translation layer now, since the pool only re-zeroes memory the VM saw written (all production writes go through translation). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
- perf_bench_test.go: synthetic ALU / load-store / call loops and interpreter setup+teardown - loader/token_perf_bench_test.go: real SPL Token Transfer through the loader/verifier/interpreter with sealevel-equivalent syscalls, in the aligned and VASA input layouts - perf_differential_test.go: deterministic random program corpus; run on two builds with SBPF_DIFF_OUT=<file> and diff the outputs; SBPF_CHECK_POOL_ZERO=1 asserts pooled buffers come back zeroed Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
…TowerSync) Measured through ExecutionCtx.ProcessInstruction so instruction-context push/pop, lamport-sum checks and timing metrics are included; each has a NoTiming variant (SkipTimingMetrics) to quantify instrumentation cost, plus a vote-state (de)serialization round trip. NOTE: written without a local build of pkg/sealevel (sandbox cannot fetch its dependencies); expect to fix compile errors on first run. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
…en transfer) Consensus-neutral performance changes to pkg/sbpf, validated against the unmodified interpreter with a 100k-program differential corpus (identical return values, errors/PCs, CU consumed, meter remaining, memory contents, input-region state) plus the package's unit tests: - meter instructions with a local due/budget pair synced around syscalls and on exit (Agave's due_insn_count scheme) instead of calling ComputeMeter.Consume per instruction - move cold opcodes to executeCold so Run drops below the compiler's "big function" threshold and Consume/Read*/Push/Pop/fast paths inline - zero only the dirty range of the pooled stack/heap in Finish (page bitmap on the fast path, byte range on the translate path) instead of 256 KiB + HeapMax per execution - per-window fast-path address translation table (Agave aligned mapping layout, branch-free v0 frame gaps, one-entry cache for VASA input regions) - 16-wide register file (no bounds checks on r[dst]/r[src]), in-place call-frame Push/Pop, precomputed internal call targets per Program pooling_test writes through the VM's translation layer now, since the pool only re-zeroes memory the VM saw written (all production writes go through translation). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
- perf_bench_test.go: synthetic ALU / load-store / call loops and interpreter setup+teardown - loader/token_perf_bench_test.go: real SPL Token Transfer through the loader/verifier/interpreter with sealevel-equivalent syscalls, in the aligned and VASA input layouts - perf_differential_test.go: deterministic random program corpus; run on two builds with SBPF_DIFF_OUT=<file> and diff the outputs; SBPF_CHECK_POOL_ZERO=1 asserts pooled buffers come back zeroed Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
…TowerSync) Measured through ExecutionCtx.ProcessInstruction so instruction-context push/pop, lamport-sum checks and timing metrics are included; each has a NoTiming variant (SkipTimingMetrics) to quantify instrumentation cost, plus a vote-state (de)serialization round trip. NOTE: written without a local build of pkg/sealevel (sandbox cannot fetch its dependencies); expect to fix compile errors on first run. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
…e loops sol_memcpy_/sol_memmove_ read the source into a fresh heap buffer and wrote it back; the copy now goes directly between the two translated slices with Go's memmove-semantics copy (overlap handled, source translated first so error precedence is unchanged, and a copy-on-write/growth of the destination region still reads the pre-write bytes because the source slice keeps the previous backing buffer alive). sol_memcmp_ uses bytes.Equal for the common equal case and word-skips to the first differing byte otherwise; sol_memset_ uses clear for zero and a doubling copy for other values. An SPL Token transfer issues two memcpy and four memcmp calls, so this is a small, allocation-free win rather than a large one. Tests: memcmpResult against the previous byte loop on 100k random inputs, memsetBytes over sizes and values, and VM-level memmove/memcpy overlap, error-ordering, copy-on-write-region and memcmp/memset checks. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
LtHash.MixIn/MixOut are 1024-lane uint16 add/subtract loops that run twice per modified account in the accounts delta hash (once for the old value, once for the new). The scalar loop costs ~560 ns per call in the sandbox and roughly 350 ns on Zen 5, so a block with ~10k modified accounts spends several milliseconds of worker CPU on lane arithmetic alone. On amd64 with AVX2 the lanes are now mixed with VPADDW/VPSUBW, 16 lanes per instruction, four vectors per iteration, unaligned loads and stores (28 ns per call here, 20x). Dispatch is a package variable set from cpu.X86.HasAVX2 (golang.org/x/sys is already a direct dependency); other architectures, CPUs without AVX2 and the purego build tag keep the portable loops, which remain the reference. Equals now compares the two arrays directly (runtime memequal) instead of a lane loop. Tests compare the assembly and the dispatched functions against the portable loops on random lanes including wrap-around values, check that MixOut inverts MixIn, that aliased operands behave, and that the generic fallback is selectable; go vet's asmdecl check passes and the package builds under -tags purego and GOARCH=arm64. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ctTQDHudYoF3FhmgmvY2y
Preserve the exact c1f7134 source tree. The new prerequisite stack contains the earlier code plus upstream epoch fixes and execution review corrections already present here; this merge changes review ancestry only.
7layermagik
force-pushed
the
7layer/execution-performance
branch
from
September 24, 2026 05:37
52a1aad to
a317eb6
Compare
This was referenced Sep 24, 2026
7layermagik
changed the title
This was referenced Sep 24, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reduce replay work after complete block assembly by combining prepared status publication, deferred checkpoints and execution during shred arrival. Include the epoch-recovery correctness fixes required by this replay stack.
Combines the scopes of #286, #289 and #292. Based on execution #281, whose integration base contains Turbine/production #287 and voting #288. Those prerequisite changes are excluded from this diff. Retarget after prerequisites merge. Kept draft for combined replay, cancellation and recovery review.
Validation: this consolidation is source-identical to
4df83876(andc1f71346); only ancestry changed. Replay/rewards, epoch fixtures and prerequisite suites passed during regrouping. The consolidation adds no deployment or new live performance claim and does not repeat complete end-to-end streaming validation. Historical checkpoint encoding improved from roughly 195 to 85 ms at the default cadence; that measures encoding work, not voting latency.